CIO Magazine
Oct 1, 2004
Decision Evolution
Automated systems are helping businesses make decisions more productively
and consistently. But they're also making a lot of entry-level jobs obsolete.
Executives had better be prepared to manage the transition.
By: Tom Davenport
Have you ever known a family in which the child went well beyond the parents?
One in which the parents didn't seem to have a lot on the ball, but they
bequeathed just enough capabilities to their child for him or her to take off?
That's just what happened in the world of automated decision making. The
parents--artificial intelligence (AI) and decision support systems (DSS)--were
ultimately disappointing despite lots of favorable hype. AI and expert systems
required those pesky knowledge engineers to create them, and they were very
difficult to maintain. Decision support systems also never really flourished,
despite being the darling of academics for decades, perhaps because they
required too much statistical expertise and too much human analysis for these
lean times.
But their offspring, a technology called automated decision systems, is taking off,
and it embodies the best attributes of each parent. Automated decision systems
are rules-based like expert systems. And like DSS, they often involve statistical
or algorithmic analysis of data. They typically make decisions in real-time after
weighing all the data and rules for a particular customer or case. Sometimes they
also carry genes from another ancestor, business process management or
workflow, leading some observers to classify them as "smart BPM" systems.
Their most salient characteristic is that they actually make a decision: what price
to charge a particular customer, whether to grant a loan or an insurance policy,
which delivery truck should be rerouted, what drug to prescribe to a diabetic
patient. In many cases their decisions are made without any human intervention
at all; in others--sometimes for legal or ethical reasons--they work alongside a
human expert such as a doctor. For the most part, these systems are being used
for decisions that must be made frequently and very rapidly using information
available online. The decision domains are relatively highly structured, with wellunderstood decision factors.
In short, this is just what knowledge workers don't need--another reason why
they're no longer necessary. If automated decision applications grow at the rate I
believe they will, there will be far fewer entry-level jobs requiring judgment and
computation. It's even easier and cheaper to automate these jobs than it is to
send them offshore, and more of an issue from a job-loss standpoint. Economists
say that productivity gains are largely responsible for 2.7 million U.S. jobs lost
since 2001, as opposed to the roughly 300,000 lost through offshoring over the
past three years. I've even heard of one company that is using Indian
programmers and statisticians to develop its automated decision systems--a
double whammy for the American knowledge worker.
These systems are being used across a variety of industries and decision types,
as my former Accenture colleague Jeanne Harris and I found after a year of
research. They are more real-time, more complex and much more pervasive than
the original versions: yield management systems in airlines that made seat
pricing decisions in the early '80s. In airlines, for example, the technology has
now been extended to a variety of operations issues, including flight scheduling
and crew and airport staff scheduling. Yield management is also being combined
with loyalty management applications to determine real-time pricing for hotel
rooms. Casino operator Harrah's Entertainment, for example, makes several
million dollars a month in incremental revenue by optimizing room rates for its
hotels and offering different rates to different levels of members in its loyalty
programs.
Although automated decision systems are common in travel and transportation,
financial services is using this technology to the greatest degree. In banking,
real-time mortgages and secured lending decisions are becoming common. For
example, LendingTree.com uses automated decision making for two purposes:
first, to decide which of its participating banks are most likely to issue a mortgage
to a customer. Second, to actually offer four mortgage deals within a few minutes
to a customer, using the banks' technology or its own.
DeepGreen Bank was designed from the ground up around automated decision
technology for home equity loans. Customers can complete an application
(including reported income) within five minutes, and an automated process
swings into action. The system pulls a credit report, invokes a scoring algorithm,
accesses an online valuation of the property, checks fraud and flood insurance,
and then makes a loan decision. About 80 percent of the time the customer
receives a final decision within two minutes. The system automatically selects a
local notary, and the customer completes the process by choosing a closing
date.
One of the first industries to adopt automated decision-making approaches was
insurance, which has been exploring automated underwriting applications for
over a decade. The industry was also an early adopter of artificial intelligence
and expert systems, but those efforts (mostly in the 1980s) did not pay off. Now,
however, the use of rules-based technology in underwriting has become
pervasive in large insurers' personal-lines businesses, and it is penetrating the
more complex processes of small-business underwriting. There have also been
some efforts to automate decision making in other insurance processes,
including claims, but adoption in these areas has been slower.
Other industries are using automated decision making to do a variety of tasks,
including: Intelligent ordering and treatment protocols in health care, Pricing and
matching of demand and supply in electronics and automobiles, Supply chain
optimization, Call center scripting, Extension of business credit and financing,
Review of credit portfolios, IT system configuration.
Technology Environment
Although automated decision making is becoming a mainstream business
activity, the technologies and vendors that support it are still relatively small
scale. Most organizations use "business rules" technologies that are available
from a few small vendors, such as Ilog, Fair Isaac and PegaSystems. Some
industries, such as insurance, have industry-specific packages available for tasks
such as underwriting, which can shorten the time to implementation. Most
organizations, however, develop their own custom systems. The technologies
involved include a rule language, a rule builder and development environment, a
rule editor and a rule repository. Rules are often embedded within other
applications, such as an insurance underwriting workflow system.
One major difference between current technology and its AI parent is that end
users or managers generally maintain the rule base, rather than technologists. In
insurance, for example, IT people and consultants often work with underwriting
experts to create an initial rule base, but ongoing management and creation of
new rules is typically done by the underwriters themselves. The goal is not just to
establish a rule base but to optimize it. If the system has been operating for some
time and there is sufficient data, it's possible to understand the contribution of
each individual rule to profitability or some operational measure and then
optimize the system by taking out rules that don't add value.
Management Issues
One of the greatest challenges for managers of these systems is finding a
sufficient number of experts to both create the system (this typically requires
several full-time experts for several months) and optimize it. One manager of a
project for small business insurance, for example, had to unplug systems for
some policy lines because he didn't have sufficient experts on staff to tune the
system and optimize the rules. He could use either underwriting experts or
actuaries for this purpose, but neither were available in sufficient numbers.
While experts may be in high demand and short supply, lower-level contributors
aren't as necessary anymore. Although we heard of no large-scale layoffs
resulting from these systems in our research, one might assume such positions
will be filled less often in the future. Managers need to think about how experts
will be developed over time without lower-level jobs to feed these positions.
When experts retire or leave the company, where will the next set of experts
come from? From a change management standpoint, managers need to
communicate as early as possible what will happen to jobs, how decision making
will be automated and centralized, and what decisions people will make versus
the system. Even experts may have issues with the change, since they'll deal
with exceptions and hard cases all the time, rather than just occasionally.
This brave new world has been a long time in coming, but it is clearly upon us
now. Businesses need to incorporate automated decision making into their
strategies and processes or they won't be successful for long. There is simply too
much data, and too many decisions to be made on it, for organizations to pass
on this technology. Some jobs may be lost, but firms that improve their
productivity in this manner will at least remain in business.
Tom Davenport is professor of IT and management at Babson College and an
Accenture Fellow. You can reach him at tdavenport@babson.edu.